Gaia's brightest very metal-poor (VMP) stars: A metallicity catalogue of a thousand VMP stars from Gaia RVS spectra

Context. Gaia DR3 has offered the scientific community a remarkable dataset of approximately one million spectra acquired with the Radial Velocity Spectrometer (RVS) in the Calcium II triplet region, that is well-suited to identify very metal-poor (VMP) stars. However, over 40% of these spectra have no released parameters by Gaia's GSP Spec pipeline in the domain of VMP stars, whereas VMP stars are key tracers of early Galactic evolution. Aims. We aim to provide spectroscopic metallicities for VMP stars using Gaia RVS spectra, thereby producing a catalogue of bright VMP stars distributed over the full sky that can serve as the basis to study early chemical evolution throughout the Galaxy. Methods. We select VMP stars using photometric metallicities from the literature and analyse the Gaia RVS spectra to infer spectroscopic metallicities for these stars. Results. The inferred metallicities agree very well with literature high-resolution metallicities with a median systematic offset of 0.1 dex and standard deviation of $\sim$0.15 dex. The purity of this sample in the VMP regime is $\sim$80% with outliers representing a mere $\sim$3%. Conclusions. We make available an all-sky catalogue of $\sim$1500 stars with reliable spectroscopic metallicities down to [Fe/H]$\sim$-4.0, of which $\sim$1000 are VMP stars. More than 75% of these stars have either no metallicity value in the literature to date or are flagged to be unreliable in their literature metallicity estimates. This catalogue of bright (G<13) VMP stars is three times larger than the current sample of well-studied VMP stars in the literature in this magnitude range, making it ideal for high-resolution spectroscopic follow-up and to study the properties of VMP stars in different parts of our Galaxy.


Introduction
The study of very metal-poor stars (VMP, [Fe/H]<-2, ie., onehundredth of solar metallicity) holds profound implications on several astrophysical processes as they provide a direct link to the early universe (e.g., Beers & Christlieb 2005).They are thought to carry the imprint of the first supernovae (see e.g., Ishigaki et al. 2018), provide us with relics of the smallest and earliest galaxies (e.g., Chiba & Beers 2000;Yuan et al. 2020;Brauer et al. 2023), and offer essential indications regarding the characteristics of the first stars and their mass distribution (see for a recent review Klessen & Glover 2023).Bright VMP stars are of special interest to the community, because we can obtain high signal-to-noise spectra and unravel in detail their chemical composition.However, finding many of these bright, very and extremely metal-poor (VMP and EMP, [Fe/H]<-3) stars is chal-lenging because they are rare among the more metal-rich and young populations of the Galaxy (Venn et al. 2004;Youakim et al. 2017;Yong et al. 2021;Bonifacio et al. 2021).
In the past decades, the Galactic archaeology community have striven to find more of these interesting and rare candidates, predominantly using the following three techniques: (i) mining large coverage spectroscopic surveys such as LAMOST (Zhao et al. 2006), SDSS (York et al. 2000), RAVE (Steinmetz et al. 2006), Gaia-ESO survey (Gilmore et al. 2012), APOGEE (Majewski et al. 2017), andGALAH (De Silva et al. 2015) (ii) prism or narrow-band photometric surveys looking at the metallicity sensitive Calcium H and K lines region as in the HK Survey (Beers et al. 1992) and the Pristine survey (Starkenburg et al. 2017;Martin, Starkenburg et al. 2023) in the north and the Hamburg-ESO Survey (Christlieb et al. 2008) and SkyMapper survey (Wolf et al. 2018) in the south, along with more recent similar methods from S-PLUS (Almeida-Fernandes et al. 2022), J-PLUS (Cenarro et al. 2019), and J-PAS (Benitez et al. 2014) surveys, (iii) using a mix of optical and infrared broad-bands to identify VMP candidates through their lack of molecular absorption near 4.6 microns (Schlaufman & Casey 2014).
As neither of these methods typically provide detailed highresolution information sensitive to the most metal-poor regime, they are almost always combined with dedicated follow-up efforts (e.g., Aguado et al. 2019;Yong et al. 2021;Li et al. 2022) to provide accurate metallicities and abundance patterns for stars pre-selected.
In this work, we combine photometric and spectroscopic efforts using some of the best datasets released by the Gaia consortium in June 2022 for the purpose of finding and characterising VMP stars.The staggering Gaia Data Release 3 (Gaia Collaboration et al. 2023, DR3) released low-resolution spectra for about 220 million stars up to a magnitude of G∼17.65 (De Angeli et al. 2022;Andrae et al. 2023b) and medium-resolution spectra for about one million stars up to a magnitude of G∼13 around the Calcium Triplet region (Recio-Blanco et al. 2023), making these spectra well-suited for the analysis of VMP stars down to metallicities of [Fe/H] ∼ -4.0 (Starkenburg et al. 2010;Carrera et al. 2013).While the former dataset has been used to provide photometric metallicities (Andrae et al. 2023a,b;Zhang et al. 2023;Yao et al. 2023;Martin, Starkenburg et al. 2023), the latter has been used by the Gaia consortium to provide the largest, and the first space-based dataset of stellar chemo-physical parameters (Recio-Blanco et al. 2023).We further leverage both catalogues in this work to analyse VMP stars specifically by preselecting them using photometric metallicities.In doing so, we end up with about one thousand VMP stars confirmed by spectroscopic metallicities that are bright and spread over the entire sky.We make this catalogue available to the community for highresolution follow-up and multiple science cases.
In Section 2 we describe the method used to select and analyse these VMP stars, Section 3 presents results and comparison of our metallicities with literature, and Section 4 summarises the properties of our VMP catalogue.

Methods
Approximately one million epoch-averaged RVS spectra were released by the Gaia consortium (Gaia Collaboration et al. 2023), with the majority of them having undergone analysis and publication during Gaia DR3.The analysis was performed using the General Stellar Parametriser for spectroscopy (GSP-Spec) module of the Astrophysical parameters inference system (Apsis) as described in Recio-Blanco et al. (2023).However, a significant fraction of the spectra were not included in the DR3 publication of stars with GSP-Spec parameters.Here, we leverage the full DR3 sample of spectra with the aim to look for metal-poor stars.In the following subsections, we will explore several approaches for identifying potential VMP targets.We accomplish this by utilizing photometric metallicities, selecting those with available Gaia RVS spectra.The main reason to use photometric pre-selection is to avoid reanalysing one million spectra, most of which are metal-rich (see Arentsen et al. 2023, for a similar approach using the LAMOST and Pristine surveys).We then summarise the spectral analysis for the RVS spectra to infer metallicities.The number of stars selected in each catalogue is shown in braces in their legend labels.

Pre-selection of potential VMP stars
To select VMP targets from the one million RVS spectra released, we use information from photometric metallicities inferred using published Gaia XP (BP-RP) spectra using the XG-BOOST algorithm by Andrae et al. (2023a, hereafter A23) and Pristine survey model by Martin, Starkenburg et al. (2023, hereafter MS23).In addition to this, we also use photometric metallicities using narrow band CaHK measurements in the northern hemisphere from the Pristine Survey data release 1 (DR1) released by MS23.We choose these three catalogues for our pre-selection due to their reliability in metallicity measurements down to the VMP regime for red giants (Aguado et al. 2019;Venn et al. 2020;Kielty et al. 2021;Lucchesi et al. 2022;Martin, Starkenburg et al. 2023).

Photometric metallicities using the XGBoost algorithm
A23 published data-driven estimates of photometric metallicities for approximately 175 million stars using low-resolution XP spectra from Gaia DR3.They train an XGBoost algorithm on APOGEE stellar parameters, supplemented with VMP stars from Li et al. (2022), utilizing various data features including spectral coefficients, narrowband fluxes, broadband magnitudes, CatWISE magnitudes, and parallax.
We select stars with photometric metallicities less than -2.0 (mh_xgboost<-2.0)from Table 2 of A23, which consists of 17 million bright red giants (G<16) with precise and pure metallicity measurements.We perform extinction correction for these stars using the method described and performed on a reduced proper motion halo catalogue by Viswanathan et al. (2023).The extinction correction involves calculating the "extinction fraction" based on the dust density model, scaling the Schlegel et al. (1998) 2D dust maps, and applying an extinction curve.These corrections allow us to calculate the amount of foreground dust for each star as a function of its parallax and location on the sky.Extinction corrected magnitudes are necessary to infer metallicities as discussed later in this section.

Photometric metallicities using synthetic CaHK ran through the Pristine survey model
Using Gaia DR3 XP spectra data, MS23 calculated synthetic CaHK magnitudes for approximately 219 million stars.These synthetic magnitudes combined with broadband Gaia information are pushed through the Pristine survey model to yield photometric metallicities.We apply the following recommended cuts on the Pristine Gaia synthetic catalogue released by MS23: -Photometric metallicity [Fe/H] less than -2.0 dex (FeHphot_CaHKsyn<-2.0 dex) -Fraction of Monte Carlo iterations used to determine [Fe/H] uncertainties is greater than 0.8 (mcfrac_CaHKsyn>0.8)-Photometric metallicity uncertainty less than 0.3 dex (0.5*(FeH_CaHKsyn_84th -FeH_CaHKsyn_16th)<0.3 dex) -Probability of being a variable star being less than 30% Pvar<0.3-Extinction on B-V magnitude is less than 0.5 (E(BV)<0.5)-Photometric quality cut that is defined as C * <σ C * (abs(Cstar)<Cstar_1sigma)

Photometric metallicities using Pristine survey DR1
The Pristine data release 1 (DR1) comes with metallicities calculated using Pristine CaHK narrow band and Gaia broad band magnitudes for all the Gaia stars with released XP spectra within the Pristine survey footprint.We use the same cuts as for Pristine Gaia synthetic catalogue for the Pristine DR1 catalogue to select potential VMP stars.The Pristine DR1 selection adds few stars that are VMP candidates and have low quality synthetic CaHK magnitudes.

Final selection
For all three photometric metallicity catalogues, we use a Gaia astrometric quality cut RUWE<1.4and parallax cut parallax_over_error>5.Hot targets and dwarfs are removed using shifted VMP PARSEC1 isochrones (Marigo et al. 2017) as shown in Figure 1.We keep the isochrone selection as wide as possible to pick up as many VMP candidate stars as possi-ble at the risk of picking up metal-rich contaminants.The final sample of stars to be analysed using the published Gaia RVS spectra are plotted as yellow, red and pink star symbols for the respective photometric catalogues they belong to.In the main VMP red giant branch, we see a large overlap between the different catalogues as expected.These duplicates are shown in purple.This comparison with all three catalogues of photometric metallicities and analysis of various stricter cuts on the Pristine catalogues and their effect on the CaMD is summarised later in Appendix B. We note that most of the relatively metal-rich outliers (that have inferred spectroscopic metallicities that are much larger than the photometric estimates) in the sample are outliers in the CaMD (Colour-absolute Magnitude Diagram).Our final sample has 1014 stars from the (A23) catalogue, 1012 stars and 149 stars from the Pristine Gaia synthetic and Pristine DR1 catalogues from (MS23) respectively.About 676 of them exist in at least two of the three catalogues.

Spectral analysis pipeline
The Gaia consortium released RVS spectra normalised for continuum.We renormalise the spectra using a spline representation as described in Appendix A to avoid systematic offsets seen when validated with high-resolution metallicities.We end up with 1441 stars that are potential VMP candidates for which we analyse the Gaia RVS spectra to infer metallicities.
For the next step, we utilize a modified pipeline based on Longeard et al. (2022) to fit equivalent widths to all three Calcium Triplet lines i.e., 850.04 nm, 854.44 nm and 866.45 nm at the same time and also compute additional radial velocity offsets in the process.Initially, we create smoothed spectra by applying a Gaussian kernel and focus on the Calcium Triplet (CaT) lines.These lines are modeled using Voigt profiles, and their positions are determined by minimizing the difference between a simulated spectrum containing only the CaT lines and the observed spectrum.The initial radial velocity estimate is obtained through cross-correlation and is typically close to zero as the publicly available Gaia RVS spectra are already in the rest frame.However owing to the handful of stars with radial velocity offsets greater than 5 km/s, we also release the radial velocity offset and error on this parameter as a part of this catalogue.A Markov Chain Monte Carlo (MCMC) algorithm is then employed to fit the observed spectra, with the aim of deriving the radial velocity offset, depth and full width at half maximum and eventually equivalent widths for the three Calcium Triplet lines.Constraints are applied to ensure the relative depths and widths of the lines are consistent such as the depth of the first line should be smaller than the second and the third which is in turn smaller than the second.The MCMC analysis is performed for each star, and the best-fit values are determined based on the maximum likelihood.
The equivalent widths (EWs) of the CaT lines are converted into metallicity measurements using the calibration provided by Carrera et al. (2013).This calibration works very well for VMP red giant stars and -due to its empirical nature -is equivalent to NLTE analyses.It requires magnitudes, calcium triplet equivalent widths and distances (inverted parallax) or height above or below the horizontal branch as inputs.To go from Gaia G magnitude to Johnston-Cross V or I magnitudes, we use the conversion defined by Riello et al. 2021.Uncertainties associated with the metallicity measurements are determined through a Monte Carlo procedure that takes into account the uncertainties in the equivalent widths, photometry, colour, distance (i.e., parallax uncertainties), and calibration relation which are the input parameters in converting EWs to metallicities.The resulting probability dis- tribution function (PDF) captures the uncertainty in the metallicity determination, with the standard deviation (using a Gaussian approximation) representing the uncertainty on the metallicities.Figure 2 illustrates some typical spectra obtained and published by Gaia RVS after renormalisation as a black line together with the best fit Voigt profile by the pipeline described above as a purple dashed line.This subsample is chosen based on a mix of signal-to-noise ratios and inferred metallicities of our sample.The median signal-to-noise ratio of our sample is 36.2 while the lowest and highest are 15.3 and 278.4 respectively.About 70% of the stars have signal-to-noise less than 50 which makes our MCMC fitting method robust on these spectra.

Comparison with other Gaia based spectroscopic metallicity catalogues
About 71.20% of the entire Gaia RVS spectra, 5.6 million stars, have been analysed by the GSP-Spec module.This percentage decreases as the metallicity decreases (down to 40% in the VMP end) and the Gaia RVS spectra are only made publicly available for a small subset of well-behaved objects (about 12.70% of the entire GSP-Spec objects).We do not find any significant trend between the fraction of un-analysed stars -or stars with bad solutions -and the signal-to-noise of the Gaia RVS spectra.About 78% of the 5.6 million GSP-spec analysed stars have the most reliable metallicities after a very strict filtering (see Figure 26 in Recio-Blanco et al. 2023).Due to the restricted wavelength range and lack of spectral information, metallicity estimates for VMP stars analyzed by the GSP-Spec module suffer from parameter degeneracy and exhibit large measurement uncertainties and systematic offsets (Kordopatis et al. 2011).Additionally, recommended quality cuts filter out a significant portion of these stars due to confusion with hot stars or challenges posed by cool K and M-type giants.
To address this, Matsuno et al. (2022, hereafter M22) aimed to break the parameter degeneracy from lack of spectral information by incorporating photometric and astrometric information and reanalyzing FGK-type stars in the GSP-Spec catalog.This approach resulted in more precise metallicity estimates, reducing uncertainties and improving agreement with high-resolution literature values.The inclusion of photometric information proved valuable in overcoming the challenges posed by lack of spectral information for VMP stars.
Because the GSP-Spec module and the catalogue by M22 uses the same spectra (directly or indirectly) to obtain metallicity and atmospheric parameters, we compare those metallicities to our inferred metallicities.Figure 3 shows a comparison of our inferred metallicities with GSP-Spec metallicities and M22 metallicities.In the left panel, big symbols for stars that pass the recommended quality cuts by Recio-Blanco et al. 2023 and small star symbols are all the stars with published GSP-Spec parameters.In the right panel, big symbols are for stars that pass the recommended GSP-Spec quality cuts as mentioned previously (defined as MP filters by M22) and small symbols are the stars with relaxed criteria improved by adding photometric information (defined as RMP filters by M22).Both catalogues have relatively small median metallicity offsets for stars that pass the quality cuts recommended.Meanwhile the GSP-Spec catalogue has higher dispersion compared to the metallicities from M22.The sample size in overlap between these catalogues is small (∼26%).This is expected due to the parameter degeneracy from lack of spectral information and thus higher uncertainties in the inferred metallicities at the VMP end.

Comparison with spectroscopic surveys and high-resolution VMP catalogues
Next, we validate our metallicities and examine the number of relatively metal-rich outliers using comparisons with spectroscopic surveys such as APOGEE DR17 (Majewski et al. 2017;Wilson et al. 2019;Abdurro'uf et al. 2022), and GALAH DR3 (Buder et al. 2021).We remove stars with STAR_BAD or FE_H_FLAG flagged from the APOGEE sample and those with flag_sp 0 or flag_fe_h 0 from the GALAH sample.The top panels of Figure 4 shows a comparison of our inferred metallicities with the existing spectroscopic surveys along with the coressponding median offset (µ 50 ) and one-sigma standard deviation (σ, using a Gaussian approximation) at the top of each panel.We can see the comparison with GALAH DR3 in the top left and APOGEE DR17 in the top right panels.The comparison with APOGEE stops at -2.5 dex which is up to where the ASPCAP pipeline assigns metallicities.In the comparison with GALAH, we see a few outliers especially in the EMP end which could be due to largely featureless GALAH spectra for EMP stars and because the GALAH DR3 pipeline is not tailored towards EMP stars with no/weak metal lines (Hughes et al. 2022).However, we visually inspected the fits and MCMC chains for these stars, and are confident in the metallicities we assign for them.The comparison with both APOGEE and GALAH shows the robustness of our metallicities in all metallicity regimes, including few metal-rich stars that we picked up as outliers in the photometric selection.The agreement with APOGEE (median offset of -0.04) is better than with GALAH (median offset of -0.08).Part of this offset might also be due to NLTE versus LTE analyses (our method provides metallicities close to the NLTE analyses).Nevertheless, there are no catastrophic outliers given the width of the distributions (σ∼0.2).The comparison with these existing spectroscopic surveys show that our inferred metallicities are very reliable especially in the VMP regime.Less than 3% of the stars in low metallicity regime ([Fe/H] T his work <-1.5) are outliers ([Fe/H] GALAH or [Fe/H] APOGEE >-1.5).
The precision in our metallicity determinations is examined by comparison with results in four homogeneously analysed high-resolution spectroscopic follow-up catalogues from Yong et al. (2021), Li et al. (2022), Hansen et al. (2018), Sakari et al. (2018).Comparison with these four catalogues allow us to study the precision of our metallicities at the VMP end and also quantify the systematic error in our method because these are large sets of homogeneously analysed stars, as opposed to assembling individual follow-up from various sources in the literature.From this comparison (see Figure 4 bottom left panel), we can see that our metallicities are accurate down to the EMP regime with median offsets as low as 0.07 dex with the largest crossmatch that we have with the Sakari et al. ( 2018) catalogue of metal-poor stars.This offset might also be due to NLTE versus LTE analyses as mentioned previously.We also compare our metallicities with the SAGA database of metal-poor stars in the literature.It is noteworthy that three of our four catalogues from the previous comparison are already a part of the SAGA database -except the Li et al. (2022) catalogue.However, this helps us quantify the precision of our metallicities collectively with a large number of stars from the literature.The comparison with the SAGA database shows that our metallicities are robust in the VMP regime and are ready to be used for an all-sky study of metalpoor stars and ideal for high-resolution spectroscopic follow-up given their relative brightness.From these comparisons, we recommend using 0.1 dex as systematic uncertainties in our metallicities.A similar comparison from the point of view of metal- licity difference and how much the offset has improved over the GSP-Spec released metallicities is presented in Appendix C.

Properties of the Gaia RVS VMPs catalogue
The VMP catalogue presented here has a brightness range of 6th to 13th magnitudes in the G band.Our catalogue of VMP stars is three times bigger than the number of VMP stars known in the literature in this brightness range (based on a comparison with the SAGA database).The median measurement uncertainty in spectroscopic metallicities inferred in our catalogue is ∼0.05 dex.The top panel of Figure 5 shows the distribution of G RVS magnitudes for the whole catalogue and different subsets such as for metal-poor (MP, [Fe/H]<-1), VMP and EMP stars.We also show the distribution of G magnitudes for a subset of stars that have a released and reliable GSP-Spec analysis.From this distribution, we can see that, typically, the released, but previously un-analysed, RVS spectra go up to a magnitude fainter (G∼13) than available GSP-Spec analysed stars (up to G∼12).Over 40% of the stars analysed in this work have no published GSP-Spec analysis and more than 75% of them do not pass the quality cuts recommended by Recio-Blanco et al. (2023) (as described for their Figure 26) for the usage of the parameters published by GSP-Spec analysis.The re-analysis of M22 of GSP-Spec results adds more reliable metallicities in the VMP regime by breaking the temperature degeneracy from lack of spectral information with external (non-Gaia) data and reduces this fraction to about 65%.This re-analysis evidently performs better at the VMP end.However, because 80% of their stars don't have a released RVS spectra, they can't visually check their spectral fits which might make it unreliable in some special cases.This is not a limitation in our method.The bottom panel of Figure 5, clearly demonstrates the large numbers of VMP stars with reliable metallicities that are added to the available datasets from GSP-Spec analysis (along with the quality cuts) and the re-analysis done by M22.It is important to note that the clump of EMP stars from GSP-Spec analysis are unreliable and disappear in the histogram with recommended quality cuts.

Summary and Outlook
With the recent Gaia Data Release 3, the Gaia consortium released about one million spectra obtained by the Radial Velocity Spectrometer (RVS) instrument.In this paper, we use these publicly available Gaia RVS spectra, with the main objective to provide an all-sky catalogue of bright VMP stars.To select potential VMP stars, we use publicly available photometric metallicities catalogues.
We present reliable metallicities for 1374 MP ([Fe/H]<-1.0),973 VMP ([Fe/H]<-2.0), and 22 EMP ([Fe/H]<-3.0)stars.We recommend to add to the reported measurement uncertainty (only ∼0.05 in the median), a systematic uncertainty of 0.1 dex derived from comparison with high-resolution analyses.This is one of the largest and only all-sky catalogue of homogenously analysed VMP stars using spectroscopy and, for the first time, using Gaia RVS spectra for a dedicated analysis of VMP stars.Our bright (6 < G RVS < 13) VMP stars catalogue increases the number of known VMP stars in this brightness range by more than a factor of three when compared to the SAGA database and is homogeneously analysed for stars over the whole sky.In our sample, over 75% of our stars has no reliable and/or no metallicities in the literature, 40% have no available parameters at all, and 93% of them have no high-resolution chemical abundances available in SAGA.
This work shows the potential of utilizing publicly accessible (archival) spectra to investigate the Galaxy's most metal-poor stars.Our catalogue is ideal for high-resolution spectroscopic follow-up due to its brightness range (meaning these stars require lower exposure times to get several other chemical abundances -few VMP stars are bright enough to be seen with a naked eye) and to study the all-sky distribution of metal-poor stars and their origin.As forthcoming Gaia data releases will unfold in the coming years with many more spectra (about a factor ten larger in DR4), we anticipate delving even deeper and substantially expanding our understanding on the origin of these very metal-poor stars in impressive numbers.

Data availability
This catalogue is made available in this temporary repository: astroakshara.github.io/rvs-paper/Gaia-RVS-VMPcatalogue-AV23b-vSep23.csv in the format shown in Table 1 before the acceptance of this paper.

Fig. 1 .
Fig. 1.Colour absolute magnitude diagram (CaMD) of stars that have photometric metallicities from MS23 (Pristine Gaia synthetic and Pristine DR1 catalogues) and A23 Table 2 giants catalogue less than -2.0 dex.Duplicates between the three catalogues are shown as purple star symbols.The selection is justified by using 13 Gyr -2.2 [M/H] PARSEC isochrone shifted by ±0.6 mag in BP 0 − RP 0 and ±0.6 mag in M G .The blue polygon is the selection area for VMP stars analysed in this paper.The number of stars selected in each catalogue is shown in braces in their legend labels.

Fig. 2 .
Fig. 2. Gaia RVS spectra of six stars in this study.The stars are presented for a range of signal-to-noise ratio and sorted by inferred spectroscopic metallicities.For clarity, the normalised flux of each star is shifted by +1.0 and the spectra between the second and the third calcium triplet line is cut-off.The calcium triplet lines are highlighted in purple and central line is indicated by gray dashed lines.

Fig. 3 .
Fig. 3. Comparison of our metallicities with other Gaia based spectroscopic metallicities such as GSP-Spec pipeline (left) and improved metallicity estimates for VMP stars in GSP-Spec catalogue by M22 (right).The number of stars (with stricter quality cuts), median (µ 50 ) and 1σ standard deviation (σ) in ∆[Fe/H] and median uncertainties on the metallicity measurements (where available) is indicated at the top and bottom of each panel.The dashed lines shows the 1:1 line and corresponding ±0.5 dex offsets.

Fig. 4 .
Fig. 4. Validation of our metallicities with existing spectroscopic surveys such as GALAH DR3 (top left), and APOGEE DR17 (top right), and high-resolution spectroscopic samples of metal-poor stars from Hansen et al. (2018); Sakari et al. (2018); Yong et al. (2021); Li et al. (2022) (bottom left), and SAGA database of VMP stars (bottom right).The number of stars, median (µ 50 ) and 1σ standard deviation (σ) in ∆[Fe/H] and median uncertainties on the metallicity measurements (where available) is indicated at the top and bottom of each panel.The dashed lines shows the 1:1 line and corresponding ±0.5 dex offsets.

Fig. 5 .
Fig. 5. Magnitude distribution of metal-poor stars in our VMP catalogue into groups of MP, VMP and EMP stars and magnitude distribution of stars with reliable metallicities from published GSP-Spec information (top).Metallicity distribution of our VMP catalogue and subsample with GSP-Spec published stars, good quality GSP-Spec stars, and reanalysed M22 VMP stars (bottom).Note that the metallicity distributions consists of metallicities inferred by each of the catalogues mentioned in the legend.

Fig. A. 1 .
Fig. A.1.Gaia RVS spectra in gray of a selected subset of 5 stars chosen with varying signal-to-noise and metallicities.Renormalised Gaia spectra in purple.The topmost spectra is the highest signal-to-noise spectra used to choose knots and masks to perform spline interpolation to infer the continuum.Gray lines represent the released Gaia RVS spectra and purple lines represent the renormalised RVS spectra using spline representation.Purple bands show the unmasked region to define the continuum and red cross symbols show the chosen knots.

Fig. B. 1 .
Fig. B.1.Comparison of our metallicities with photometric metallicities from Gaia XP based catalogues such as Pristine Gaia Synthetic catalogue from MS23 (left), Table 2 giants catalogue from A23 (right) and dedicated CaHK based metallicity survey such as Pristine DR1 from MS23 (middle).The number of stars, median (µ 50 ) and 1σ standard deviation (σ) in ∆[Fe/H] and median uncertainties on the metallicity measurements (where available) is indicated at the top of each panel.The dashed lines shows the 1:1 line and coressponding ±0.5 dex offsets.

Fig. B. 2 .
Fig. B.2. Colour absolute magnitude (CaMD) diagram of stars that have photometric metallicities from MS23 (Pristine Gaia synthetic and Pristine DR1 catalogues) less than -2.0 dex, showing the consequence of wide and strict selections, colour-coded by the derived spectroscopic metallicities.

Fig. C. 1 .
Fig. C.1.Precision of our metallicities (bottom) over GSP-Spec metallicities (top) validated using literature metallicities.The dashed lines shows the 0.0±0.5 dex metallicity difference lines.The gray shaded region in the bottom panel indicates the part of the figure where it's impossible to have data points due to the low metallicity selection ([Fe/H] T his work <-1.5).

Table 1 .
Description of the columns of the Gaia RVS spectra VMP stars catalogue made available publicly in this work (currently accessible in a temporary repository before acceptance astroakshara.github.io/rvs-paper/Gaia-RVS-VMP-catalogue-AV23b-vSep23.csv).